Gradient Flows, Heat Equation, and Brownian Motion on Time ...€¦ · Heat Flow on Metric Measure...

Gradient Flows, Heat Equation, and BrownianMotion on Time-dependent Metric Measure

Spaces

Theo Sturm

partly with Eva Kopfer, Matthias Erbar

Hausdorff Center for MathematicsInstitute for Applied Mathematics

Universitat Bonn

Outline

Heat flow as gradient flow for the entropy

Dirichlet heat flow

Heat flow on time-dependent metric measure spaces

(Super-) Ricci flows

Heat Flow on Metric Measure Spaces

(X , d) complete separable metric space, m locally finite measure

Heat equation on X

• either as gradient flow on L2(X ,m) for the energy

E(u) =1

|∇u|2 dm = lim infv→u in L2(X,m)

v∈Lip(X ,d)

(lipxv)2 dm(x)

with |∇u| = minimal weak upper gradient

• or as gradient flow on P2(X , d) for the relative entropy

Ent(u m) =

u log u dm.

Theorem (Ambrosio/Gigli/Savare).

For arbitrary metric measure spaces (X , d ,m) satisfying CD(K ,∞) both ap-proaches coincide.

Note: Heat flow Ptu not necessarily linear.

Heat Flow on Metric Measure Spaces

Heat equation on X

• either as gradient flow on L2(X ,m) for the energy

E(u) =1

|∇u|2 dm = lim infv→u in L2(X,m)

v∈Lip(X ,d)

(lipxv)2 dm(x)

with |∇u| = minimal weak upper gradient

• or as gradient flow on P2(X , d) for the relative entropy

Ent(u m) =

u log u dm.

Theorem (Ambrosio/Gigli/Savare).

For arbitrary metric measure spaces (X , d ,m) satisfying CD(K ,∞) both ap-proaches coincide.

Note: Heat flow Ptu not necessarily linear.

Synthetic Ricci Bounds for Metric Measure Spaces

Definition. The Curvature-Dimension Condition CD(K ,∞)

⇐⇒ ∀µ0, µ1 ∈ P(X ) : ∃ geodesic (µt)t s.t. ∀t ∈ [0, 1]:

S(µt) ≤ (1− t) S(µ0) + t S(µ1)− K

2t(1− t)W 2(µ0, µ1)

with Boltzmann entropy

S(µ) = Ent(µ|m) =

{ ∫Xρ log ρ dm , if µ = ρ ·m

+∞ , if µ 6� m ��

��

and Kantorovich-Wasserstein metric

W (µ0, µ1) = infq

{∫X×X

d2(x , y) d q(x , y) : (π1)∗q = µ0, (π2)∗q = µ1

Synthetic Ricci Bounds for Metric Measure Spaces

Definition. The Curvature-Dimension Condition CD(K ,∞)

⇐⇒ ∀µ0, µ1 ∈ P(X ) : ∃ geodesic (µt)t s.t. ∀t ∈ [0, 1]:

S(µt) ≤ (1− t) S(µ0) + t S(µ1)− K

2t(1− t)W 2(µ0, µ1)

or equivalently

∂tS(µt)∣∣t=1− ∂tS(µt)

∣∣t=0≥ K ·W 2(µ0, µ1)

with Boltzmann entropy

S(µ) = Ent(µ|m) =

{ ∫Xρ log ρ dm , if µ = ρ ·m

+∞ , if µ 6� m��

��

and Kantorovich-Wasserstein metric

W (µ0, µ1) = infq

{∫X×X

d2(x , y) d q(x , y) : (π1)∗q = µ0, (π2)∗q = µ1

The Relevance of Lower Ricci Bounds

Nonnegative Ricci curvature implies that – in many respects – optimaltransports, heat flows, and Brownian motions behave as nicely as on Euclideanspaces. For instance

Heat kernel comparison

pt(x , y) ≥ (4πt)−n/2 exp

(−d2(x , y)

)Li-Yau estimates

Gradient estimates|∇Ptu| ≤ Pt(|∇u|)

Transport estimatesW (Ptµ,Ptν) ≤W (µ, ν)

∀x , y : ∃ coupled Brownian motions (Xt ,Yt)t≥0 starting at (x , y) s.t.P-a.s. for all t ≥ 0

d(Xt ,Yt) ≤ d(x , y)

Indeed, Ric ≥ 0 is necessary and sufficient for each of the latter properties.

Among the applications:‘Market Fragility, Systemic Risk, and Ricci Curvature’ (Sandhu et al. 2015)‘Ricci curvature and robustness of cancer networks’ (Tannenbaum et al. 2015)

The Relevance of Lower Ricci Bounds

Nonnegative Ricci curvature implies that – in many respects – optimaltransports, heat flows, and Brownian motions behave as nicely as on Euclideanspaces. For instance

Heat kernel comparison

pt(x , y) ≥ (4πt)−n/2 exp

(−d2(x , y)

)Li-Yau estimates

Gradient estimates|∇Ptu| ≤ Pt(|∇u|)

Transport estimatesW (Ptµ,Ptν) ≤W (µ, ν)

∀x , y : ∃ coupled Brownian motions (Xt ,Yt)t≥0 starting at (x , y) s.t.P-a.s. for all t ≥ 0

d(Xt ,Yt) ≤ d(x , y)

Indeed, Ric ≥ 0 is necessary and sufficient for each of the latter properties.

Among the applications:‘Market Fragility, Systemic Risk, and Ricci Curvature’ (Sandhu et al. 2015)‘Ricci curvature and robustness of cancer networks’ (Tannenbaum et al. 2015)

The Gradient Flow Perspective

Powerful consequences of the gradient flow perspective (Otto, Villani)

HessS ≥ K

|∇S |2 ≥ 2K · S

S ≥ K/2 ·W2(·, ν∞)2

The Gradient Flow Perspective

Powerful consequences of the gradient flow perspective (Otto, Villani)

HessS ≥ K ”Bakry-Emery inequality”

|∇S |2 ≥ 2K · S ”log. Sobolev inequality”

S ≥ K/2 ·W2(·, ν∞)2 ”Talagrand inequality”

Dirichlet Heat Flow

(X , d) complete separable metric space, Y ⊂ X open with ∅ 6= Y 6= X .

Charged probability measures on X

σ = (σ+, σ−) : σ± ∈ P sub(X ), σ+|X\Y = σ−|X\Y , σ+(X ) + σ−(X ) = 1;

Effective measure σ0 = σ+ − σ−, total measure σ = σ+ + σ−.

Kantorovich-Wasserstein distance between two charged probability measuresσ = (σ+, σ−) and τ = (τ+, τ−)

W2(σ, τ)2 = inf{∫ ∫

d(x , y)2dq++(x , y) +

∫ ∫d∗(x , y)2dq+−(x , y)

∫ ∫d∗(x , y)2dq−+(x , y) +

∫ ∫d(x , y)2dq−−(x , y) :

σi = σi+ + σi−, τ j = τ+j + τ−j , qij ∈ Cpl(σij , τ ij), i , j ∈ {+,−}}

whered∗(x , y) := inf

z∈X\Y

[d(x , z) + d(z , y)

Dirichlet Heat Flow

(X , d) complete separable metric space, Y ⊂ X open with ∅ 6= Y 6= X .Charged probability measures on X

W2(σ, τ)2 = inf{∫ ∫

d(x , y)2dq++(x , y) +

∫ ∫d∗(x , y)2dq+−(x , y)

∫ ∫d∗(x , y)2dq−+(x , y) +

∫ ∫d(x , y)2dq−−(x , y) :

z∈X\Y

[d(x , z) + d(z , y)

Dirichlet Heat Flow

(X , d) complete separable metric space, Y ⊂ X open with ∅ 6= Y 6= X .Charged probability measures on X

W2(σ, τ)2 = inf{∫ ∫

d(x , y)2dq++(x , y) +

∫ ∫d∗(x , y)2dq+−(x , y)

∫ ∫d∗(x , y)2dq−+(x , y) +

∫ ∫d(x , y)2dq−−(x , y) :

z∈X\Y

[d(x , z) + d(z , y)

Dirichlet Heat Flow

Thm. (Profeta, St. ‘17) Assume X Riem. mfd. with Ric ≥ K , Y convexsubset. Then Dirichlet heat semigroup (P0

t )t>0 on Y is given by the effectivemeasure of the gradient flow for the Boltzmann entropy Ent(σ+|m)+Ent(σ−|m)within the space of charged prob. measures w.r.t. the distance W2.

Cor.1 ∀ subprobab. µ, ν on Y

W 2 (P0t µ,P

0t ν) ≤ e−Kt ·

W 2 (µ, ν)

with Kantorovich-Wasserstein distance between subprobabilities µ, ν ∈ P sub(Y )

W 2 (µ, ν) = inf{W2

((σ+, σ−), (τ+, τ−)

): σ+ − σ− = µ, τ+ − τ− = ν

}= inf

((µ+ ρ, ρ), (ν + η, η)

): ρ, η ∈ P sub(X ),

(µ+ 2ρ)(X ) = 1, (ν + 2η)(X ) = 1}.

Cor.2∣∣∇P0

t u∣∣ ≤ e−Kt · PN

t |∇u|

Dirichlet Heat Flow

W 2 (P0t µ,P

0t ν) ≤ e−Kt ·

W 2 (µ, ν)

W 2 (µ, ν) = inf{W2

((σ+, σ−), (τ+, τ−)

): σ+ − σ− = µ, τ+ − τ− = ν

}= inf

((µ+ ρ, ρ), (ν + η, η)

): ρ, η ∈ P sub(X ),

(µ+ 2ρ)(X ) = 1, (ν + 2η)(X ) = 1}.

Cor.2∣∣∇P0

t |∇u|

Dirichlet Heat Flow

W 2 (P0t µ,P

0t ν) ≤ e−Kt ·

W 2 (µ, ν)

W 2 (µ, ν) = inf{W2

((σ+, σ−), (τ+, τ−)

): σ+ − σ− = µ, τ+ − τ− = ν

}= inf

((µ+ ρ, ρ), (ν + η, η)

): ρ, η ∈ P sub(X ),

(µ+ 2ρ)(X ) = 1, (ν + 2η)(X ) = 1}.

Cor.2∣∣∇P0

t |∇u|

Heat Flow on Time-dependent MM-Spaces

Aim. Study heat flow on I × X where I = (0,T ) ⊂ R and (X , dt ,mt) is metricmeasure space (∀t ∈ I )

Many challenges.

Define/study solutions for heat equation ∂tu = ∆tu

Define/study gradient flows for energy and for entropy

Find correct time-dependent versions of Bakry-Emery (=Bochner) and ofLott-St.-Villani conditions

Establish equivalence between Eulerian and Lagrangian approach

The EVI Formulation of Gradient Flows

xt = −∇V (xt)

〈xt − z , xt〉 ≤ 〈∇V (xt), z − xt〉 (∀z)

⇓ V convex

2∂t |xt − z |2 ≤ V (z)− V (xt) (∀z)

Evolution variational inequality (in Hilbert spaces) for ‘static’ V

Powerful extension to ‘static’ metric spaces (Ambrosio, Gigli, Savare)

xt = −∇V (xt)

〈xt − z , xt〉 ≤ 〈∇V (xt), z − xt〉 (∀z)

⇓ V convex

2∂t |xt − z |2 ≤ V (z)− V (xt) (∀z)

xt = −∇V (xt)

〈xt − z , xt〉 ≤ 〈∇V (xt), z − xt〉 (∀z)

⇓ V convex

2∂t |xt − z |2 ≤ V (z)− V (xt) (∀z)

xt = −∇V (xt)

〈xt − z , xt〉 ≤ 〈∇V (xt), z − xt〉 (∀z)

⇓ V convex

2∂t |xt − z |2 ≤ V (z)− V (xt) (∀z)

Gradient Flows on Time-dependent MM-Spaces

Question: How to define gradient flow for time-dependent potentialV : I × X → (−∞,∞] on time-dependent metric space (X , dt)t∈I?

Examples:

gradient flow for Boltzmann entropy St on (P,Wt)t∈I

gradient flow for Dirichlet energy Et on L2(X ,mt)t∈I

Definition

An absolutely continuous curve (xt)t∈J will be called dynamic backward EVI−-gradient flow for V if for all t ∈ J and all z ∈ Dom(Vt)

2∂−s d2

s,t(xs , z)∣∣∣s=t−

≥ Vt(xt)− Vt(z)

ds,t(x , y) := inf

{∫ 1

|γa|2s+a(t−s)da

with infimum over all absolutely continuous curves (γa)a∈[0,1] in X from x to y .

Examples:

gradient flow for Boltzmann entropy St on (P,Wt)t∈I

gradient flow for Dirichlet energy Et on L2(X ,mt)t∈I

Definition

2∂−s d2

s,t(xs , z)∣∣∣s=t−

≥ Vt(xt)− Vt(z)

ds,t(x , y) := inf

{∫ 1

|γa|2s+a(t−s)da

Heat Flow on Time-dependent MMS

For the sequel, a 1-parameter family of metric measure spaces (X , dt ,mt),t ∈ I ⊂ R will be given s.t. ∀s, t ∈ I

the mm-space (X , dt ,mt) satisfies CD∗(K ,N) and has linear heat flow

log dt (x,y)ds (x,y)

≤ C |s − t|

mt(dx) = e−ft (x)m0(dx) for some f ∈ Lip(I × X )

Thus ∀t ∈ I :∃ Dirichlet form Et , Laplacian ∆t , squared gradient Γt(u) = |∇tu|2.

Theorem (‘Heat equation’)

∃ heat kernel p on {(t, s, x , y) ∈ I 2 × X 2 : t > s}, Holder contin-uous in all variables and satisfying the propagator property pt,r (x , z) =∫pt,s(x , y)ps,r (y , z) dms(y), such that

(t, x) 7→ pt,s(x , y) solves the heat equation ∂tut = ∆tut on (s,T )× X

(s, y) 7→ pt,s(x , y) solves the adjoint heat equation∂svs = −∆svs + (∂s fs) · vs on (0, t)× X

For the sequel, a 1-parameter family of metric measure spaces (X , dt ,mt),t ∈ I ⊂ R will be given s.t. ∀s, t ∈ I

the mm-space (X , dt ,mt) satisfies CD∗(K ,N) and has linear heat flow

log dt (x,y)ds (x,y)

≤ C |s − t|

mt(dx) = e−ft (x)m0(dx) for some f ∈ Lip(I × X )

Thus ∀t ∈ I :∃ Dirichlet form Et , Laplacian ∆t , squared gradient Γt(u) = |∇tu|2.

Theorem (‘Heat equation’)

∃ heat kernel p on {(t, s, x , y) ∈ I 2 × X 2 : t > s}, Holder contin-uous in all variables and satisfying the propagator property pt,r (x , z) =∫pt,s(x , y)ps,r (y , z) dms(y), such that

(t, x) 7→ pt,s(x , y) solves the heat equation ∂tut = ∆tut on (s,T )× X

(s, y) 7→ pt,s(x , y) solves the adjoint heat equation∂svs = −∆svs + (∂s fs) · vs on (0, t)× X

(i) ∀s ∈ I ,∀h ∈ L2(X ,ms) : ∃! solution to heat equation ∂tut = ∆tut on(s,T )× X with us = h given by

ut(x) = Pt,sh(x) :=

∫pt,s(x , y)h(y) dms(y)

(ii) ∀t ∈ I , ∀g ∈ L2(X ,mt) : ∃! solution to the adjoint heat equation∂svs = −∆svs + (∂s fs) · vs on (0, t)× X with vt = g given by

vs(y) = P∗t,sg(y) :=

∫pt,s(x , y)g(x) dmt(x)

(iii) Define dual heat flow Pt,s : P(X )→ P(X ) by

(Pt,sµ)(dy) =

[∫pt,s(x , y) dµ(x)

]ms(dy).

In particular, Pt,s

(g ·mt

)=(P∗t,sg

)·ms and∫

hd(Pt,sµ

∫ (Pt,sh

(i) ∀s ∈ I ,∀h ∈ L2(X ,ms) : ∃! solution to heat equation ∂tut = ∆tut on(s,T )× X with us = h given by

ut(x) = Pt,sh(x) :=

∫pt,s(x , y)h(y) dms(y)

(ii) ∀t ∈ I , ∀g ∈ L2(X ,mt) : ∃! solution to the adjoint heat equation∂svs = −∆svs + (∂s fs) · vs on (0, t)× X with vt = g given by

vs(y) = P∗t,sg(y) :=

∫pt,s(x , y)g(x) dmt(x)

(iii) Define dual heat flow Pt,s : P(X )→ P(X ) by

(Pt,sµ)(dy) =

[∫pt,s(x , y) dµ(x)

]ms(dy).

In particular, Pt,s

(g ·mt

)=(P∗t,sg

)·ms and∫

hd(Pt,sµ

∫ (Pt,sh

Heat Flow and its Dual as Gradient Flows

Theorem. Assume mt ≤ eL(t−s)ms .

∀u ∈ Dom(E), ∀s the heat flow t 7→ ut = Pt,su is the unique dynamical forwardEVIL-gradient flow for the energy 1

2E , that is, ∀v ∈ Dom(E), ∀t

2∂−s ‖us − v‖2s,t

∣∣s=t− +

4‖ut − v‖2t ≥

2Et(ut)−

2Et(v).

Theorem. Assume that (X , dt ,mt)t∈I is a super-Ricci flow.

∀µ ∈ Dom(S),∀T the dual heat flow s 7→ µs = PT ,sµ is the unique dy-namical backward EVI-gradient flow for the Boltzmann entropy S , that is,∀σ ∈ Dom(S), ∀t

2∂−s W 2

s,t(µs , σ)∣∣s=t− ≥ St(µt)− St(σ).

Here W 2s,t(µ0, µ1) := inf

0|µa|2s+a(t−s)da with infimum over all AC2-curves

(µa)a∈[0,1] in P(X ) connecting µ0 and µ1.

Heat Flow and its Dual as Gradient Flows

Theorem. Assume mt ≤ eL(t−s)ms .

∀u ∈ Dom(E), ∀s the heat flow t 7→ ut = Pt,su is the unique dynamical forwardEVIL-gradient flow for the energy 1

2E , that is, ∀v ∈ Dom(E), ∀t

2∂−s ‖us − v‖2s,t

∣∣s=t− +

4‖ut − v‖2t ≥

2Et(ut)−

2Et(v).

Theorem. Assume that (X , dt ,mt)t∈I is a super-Ricci flow.

∀µ ∈ Dom(S),∀T the dual heat flow s 7→ µs = PT ,sµ is the unique dy-namical backward EVI-gradient flow for the Boltzmann entropy S , that is,∀σ ∈ Dom(S), ∀t

2∂−s W 2

s,t(µs , σ)∣∣s=t− ≥ St(µt)− St(σ).

Here W 2s,t(µ0, µ1) := inf

0|µa|2s+a(t−s)da with infimum over all AC2-curves

(µa)a∈[0,1] in P(X ) connecting µ0 and µ1.

Super Ricci Flows

A family of Riemannian manifolds (M, gt), t ∈ (0,T ), is called super-Ricci flowiff

Rict +1

2∂tgt ≥ 0.

Two main examples

Static manifolds with Ric ≥ 0 (‘elliptic case’)

Ricci flows Rict = − 12∂tgt (‘minimal super-Ricci flows’)

Hamilton’81, Perelman’02, . . . , McCann/Topping’10, Lott’09,Arnaudon/Coulibaly/Thalmaier’08, Kuwada/Philipowski’11, X.-D.Li’14,Kleiner/Lott’14, Haslhofer/Naber’15

Super Ricci Flows

A family of Riemannian manifolds (M, gt), t ∈ (0,T ), is called super-Ricci flowiff

Rict +1

2∂tgt ≥ 0.

Given a 1-parameter family of metric measure spaces (X , dt ,mt), t ∈ I ⊂ R.Consider the function

S : I × P(X )→ (−∞,∞], (t, µ) 7→ St(µ) = Ent(µ|mt)

where P(X ) is equipped with the 1-parameter family of metrics Wt (=L2-Wasserstein metrics w.r.t. dt).

Definition.

(X , dt ,mt)t∈I is super-Ricci flow iff for a.e. t and every Wt-geodesic (µa)a∈[0,1]

∂aSt(µa)∣∣a=0− ∂aSt(µ

a)∣∣a=1≤ 1

2∂−t W 2

t−(µ0, µ1).

Characterization of Super-Ricci Flows

Theorem. The following are equivalent

a)∣∣a=1≤ 1

2∂tW

2t (µ0, µ1)

Ws(Pt,sµ, Pt,sν) ≤Wt(µ, ν)

∀x , y , ∀t : ∃ coupled backward Brownian motions (Xs ,Ys)s≤t starting att in (x , y) s.t. ds(Xs ,Ys) ≤ dt(x , y) a.s. for all s ≤ t

|∇t(Pt,su)|2 ≤ Pt,s(|∇su|2)

Γ2,t ≥ 12∂tΓt where Γ2,t(u) = 1

2∆t |∇tu|2 − 〈∇tu,∇t∆tu〉

a)∣∣a=1≤ 1

2∂tW

2t (µ0, µ1)

|∇t(Pt,su)|2 ≤ Pt,s(|∇su|2)

2∆t |∇tu|2 − 〈∇tu,∇t∆tu〉

a)∣∣a=1≤ 1

2∂tW

2t (µ0, µ1)

|∇t(Pt,su)|2 ≤ Pt,s(|∇su|2)

2∆t |∇tu|2 − 〈∇tu,∇t∆tu〉

a)∣∣a=1≤ 1

2∂tW

2t (µ0, µ1)

|∇t(Pt,su)|2 ≤ Pt,s(|∇su|2)

2∆t |∇tu|2 − 〈∇tu,∇t∆tu〉

a)∣∣a=1≤ 1

2∂tW

2t (µ0, µ1)

|∇t(Pt,su)|2 ≤ Pt,s(|∇su|2)

2∆t |∇tu|2 − 〈∇tu,∇t∆tu〉

Functional Inequalities for Super-Ricci Flows

Theorem. For every super-Ricci flow (X , dt ,mt)t∈I

Local Poincare inequalities

2(t − s)Γt(Pt,su) ≤ Pt,s(u2)− (Pt,su)2 ≤ 2(t − s)Pt,s(Γsu)

Local logarithmic Sobolev inequalities

(t − s)Γt(Pt,su)

Pt,su≤ Pt,s(u log u)− (Pt,su) log(Pt,su) ≤ (t − s)Pt,s

Dimension-free Harnack inequality: ∀α > 1

(Pt,su)α(y) ≤ Pt,suα(x) · exp

( αd2t (x , y)

4(α− 1)(t − s)

(t − s)Γt(Pt,su)

( αd2t (x , y)

4(α− 1)(t − s)

(t − s)Γt(Pt,su)

( αd2t (x , y)

4(α− 1)(t − s)

Characterization of Ricci Flows

(X , dt ,mt)t∈I is a Ricci flow iff super-Ricci flow and ∀x , t

limy→x

∂s logWs

(Pt,sδx , Pt,sδy

)∣∣∣s=t−

If (X , dt ,mt)t∈I is Ricci flow then ∀t, ∀dt -geodesics (γa)a∈[0,1] the length of thecurve

µas = Pt,sδγa , a ∈ [0, 1]

in (P(X ),Ws) asymptotically for s ↗ t does not change

∂sLength(µ·s)∣∣∣

s=t−= 0

Theorem.

Assume that X is N-cone of some mm-space Y . Then X is Ricci bounded if andonly if N ∈ N and X = RN+1.

limy→x

∂s logWs

(Pt,sδx , Pt,sδy

)∣∣∣s=t−

µas = Pt,sδγa , a ∈ [0, 1]

s=t−= 0

Theorem.

limy→x

∂s logWs

(Pt,sδx , Pt,sδy

)∣∣∣s=t−

µas = Pt,sδγa , a ∈ [0, 1]

s=t−= 0

Theorem.

Thank You For Your Attention!

Synthetic Upper Bounds for Ricci Curvature

Theorem

For weighted Riem with Ric∞ > −K0, for all non-conjugate x , y

Kx,y ≤ −∂−t logW(Ptδx ,Ptδy

)∣∣t=0+

≤ Kx,y + σx,y tan2 (√σx,yd(x , y)/2)

with Kx,y = average Ricci curvature along min. geodesic from x to yand σx,y = maximal modulus of sect. curv. along this geodesic.

Singular examples

Doubling of B1(0) ⊂ Rn in n ≥ 2

Cone over S2(1/√

3)× S2(1/√

Cone over circle of length α < 2π

Proposition

For the cone over circle of length α < 2π

W(Ptδx ,Ptδy

{d(x , y)−

√πt · 2

αsin α

2+ O(t), if x or y is the vertex

d(x , y) + o(t), else.

Theorem

)∣∣t=0+

Singular examples

Cone over S2(1/√

3)× S2(1/√

Proposition

W(Ptδx ,Ptδy

{d(x , y)−

√πt · 2

αsin α

Theorem

)∣∣t=0+

Singular examples

Cone over S2(1/√

3)× S2(1/√

Proposition

W(Ptδx ,Ptδy

{d(x , y)−

√πt · 2

αsin α

Definition

2∂−s d2

s,t(xs , z)∣∣∣s=t−

≥ Vt(xt)− Vt(z)

ds,t(x , y) := inf

{∫ 1

|γa|2s+a(t−s)da

Definition

(xt)t∈J is called dynamic backward EVI(K ,N)-gradient flow if ∀z , ∀t

2∂−s d2

s,t(xs , z)∣∣∣s=t− K

2· d2

t (xt , z) ≥Vt(xt)− Vt(z)

(∂aVt(γ

(1− a)da

Definition

2∂−s d2

s,t(xs , z)∣∣∣s=t−

≥ Vt(xt)− Vt(z)

ds,t(x , y) := inf

{∫ 1

|γa|2s+a(t−s)da

Definition

(xt)t∈J is called dynamic backward EVI(K ,N)-gradient flow if ∀z , ∀t

2∂−s d2

s,t(xs , z)∣∣∣s=t− K

2· d2

t (xt , z) ≥Vt(xt)− Vt(z)

(∂aVt(γ

(1− a)da

The Curvature-Dimension Condition CD∗(K ,N)

Def. A metric measure space (X , d ,m) satisfies CD∗(K ,N)

⇐⇒ S:= Ent(.) is (K,N)-convex on (P(X ),W )

⇐⇒ ∂tS(µt)∣∣t=1− ∂tS(µt)

∣∣t=0≥ K ·W 2(µ0, µ1) + 1

(S(µ0)− S(µ1)

sec ≥ 0 ⇐⇒ dist concave ric ≥ 0 ⇐⇒ vol1/n concave

Proposition. CD∗(0,N) ⇐⇒ ∀µ0, µ1 ∈ P(X ) : ∃ geodesic (µt)t s.t.

SN(µt |m) ≤ (1− t)SN(µ0|m) + t SN(µ1|m))

where SN(ν|m) = −∫Xρ1−1/N dm for ν = ρ ·m + νs .

Example: SN(ν|m) = −m(A)1/N if ν = unif. distrib. on A ⊂ X

The Curvature-Dimension Condition CD∗(K ,N)

Def. A metric measure space (X , d ,m) satisfies CD∗(K ,N)

⇐⇒ S:= Ent(.) is (K,N)-convex on (P(X ),W )

⇐⇒ ∂tS(µt)∣∣t=1− ∂tS(µt)

∣∣t=0≥ K ·W 2(µ0, µ1) + 1

(S(µ0)− S(µ1)

sec ≥ 0 ⇐⇒ dist concave ric ≥ 0 ⇐⇒ vol1/n concave

Proposition. CD∗(0,N) ⇐⇒ ∀µ0, µ1 ∈ P(X ) : ∃ geodesic (µt)t s.t.

SN(µt |m) ≤ (1− t)SN(µ0|m) + t SN(µ1|m))

where SN(ν|m) = −∫Xρ1−1/N dm for ν = ρ ·m + νs .

Example: SN(ν|m) = −m(A)1/N if ν = unif. distrib. on A ⊂ X

Gradient Flows, Heat Equation, and Brownian Motion on Time ...€¦ · Heat Flow on Metric Measure...

Documents

Fourier’s Heat Conduction Equation

Electric Circuits & The Heat Equation

Specific Heat and Using the Heat Equation Thermal Energy Thermal Energy Heat Heat Temperature Temperature

Hyperbolicity of the heat equation

Geometric Heat Equation and Nonlinear Diffusion of …shape/publications/cviu96.pdf · GEOMETRIC HEAT EQUATION AND NONLINEAR DIFFUSION 307 The curvature deformation equation (7),

HEAT CONDUCTION EQUATION H - EECL-COURSE · HEAT CONDUCTION EQUATION H eat ... 2–3 General Heat Conduction Equation 74 ... One-Dimensional Heat Conduction Problems 86 2–6 Heat

A Stochastic Heat Equation

The continuity equation on metric measure spaces

2 Heat Equation - Stanford University · 2003. 6. 22. · 2 Heat Equation 2.1 Derivation Ref: Strauss, Section 1.3. Below we provide two derivations of the heat equation, ut ¡kuxx

The 1-D Heat Equation

2012 Heat Equation - Parkland College

HEAT TRANSFER EQUATION SHEET - UTRGV Faculty Web Transfer... · 1 HEAT TRANSFER EQUATION SHEET Heat Conduction Rate Equations (Fourier's Law) Heat Flux : 𝑞. 𝑥′′ = −𝑘

2-Heat Conduction Equation

PDF 2 Heat Equation

Process Heat Transfer Lec 3: Heat Conduction Equation

Using the Revised Universal Soils Loss Equation as a Metric

The Heat Equation - users.math.msu.eduusers.math.msu.edu/users/yan/847ch3.pdf · The Heat Equation The heat equation, also known as di usion equation, ... Especially, we expect to

Chapter 1: Heat equation

Applications: Discretized Heat Equation

Heat equation & Calorimetry Key - Weebly